Bryan Leonard — Researcher, Engineer, Builder

About

Independent researcher building from zero.

I study the gap between what people know and what they can say. That question led me to build Expression-Gated Consciousness — a formal framework for measuring how expression gates transparency — and to design LOLM, a novel language model architecture from scratch.

Co-founder of Qira LLC with my brother Brandyn Leonard. Self-taught full-stack engineer. Currently training models at billion-parameter scale on Google’s TPU Research Cloud while shipping production SaaS, Web3 infrastructure, and autonomous systems.

No university. No lab. No venture capital. Just work.

Open to: Municipal technology partnerships, research collaboration, compute grants, and teams solving hard problems.

PyTorch Python React Next.js TypeScript Solana FastAPI NLP CUDA Supabase

View Resume →

The Work

Flagship Research

Expression-Gated Consciousness

A mathematical framework and live empirical study on how emotional knowledge gates conscious expression. 44+ participants. Three distinct response types confirmed by data.

N=44+ subjects · Pearson r=0.311 · Comfort gap: 5.6pts

g(K) = 4K(1−K) — Brandyn Leonard

Live

Take the study → Research page (The Gate) → Preprint on Zenodo → Blind rating tool →

Live Product — In Talks With ADOT

Phoenix Traffic Intelligence

Real-time traffic intelligence with AI-powered predictions, TomTom live speeds, and cascade failure detection. Monitoring 8 Phoenix freeway corridors 24/7 with automated crew dispatch recommendations.

Live AZ-511 data · 375K+ corridor snapshots · 8 corridors · 2-min sweep cycle

Live

Public corridor monitor → Technical dashboard → GitHub →

Original Architecture — Google TPU Grant

LOLM — Custom Language Model

Building a transformer architecture from scratch. Not fine-tuning — original architecture targeting frontier-class capability on Google TPU infrastructure.

10B–100B parameter target · Google TPU Research Cloud grant

Training

Project site → GitHub →

Codey — AI Coding Intelligence

A coding platform that understands your codebase as a connected network. Structural analysis, dependency mapping, and intelligent code generation.

SaaS · GPT-4o + 11 providers · VS Code extension · CLI

In Development

Try the beta → Landing page → GitHub →

Codey — AI Coding Agent → NFET Live Dashboard → ADOT Public Dashboard → Command Center → Claude Usage Dashboard → All GitHub Repos → EGC Rater Interface →

Live Infrastructure

Every system listed here is running in production right now. Not demos. Not mockups. Live services on real data, self-funded from Phoenix.

Live 24/7 ■

Phoenix Traffic Intelligence

Real-time predictive cascade model for I-10 and I-17 corridors. Ingests live AZ-511 data, computes Level of Service grades, and generates 2-minute AI crew dispatch recommendations for 8 Phoenix freeway corridors.

375K+ corridor snapshots · 10-second polling · AI sweep every 2 min

View live public dashboard →

Training ◆

LOLM — Custom Transformer

Original language model architecture designed from scratch. Not fine-tuning an existing model — building the architecture, tokenizer, training loop, and evaluation pipeline. Running on Google TPU Research Cloud infrastructure.

10B–100B parameter target · Google TPU Research Cloud grant

View architecture on GitHub →

Live Study ●

EGC — Consciousness Research

Live empirical psychology study with 44+ participants. Original mathematical framework (g(K) = 4K(1−K), by Brandyn Leonard) measuring how emotional knowledge gates conscious expression. Preprint on Zenodo.

N=44+ · Pearson r=0.311 · 3 confirmed response types · Peer-reviewable

Read the preprint →

In Development ◊

Codey — AI Coding Agent

Full-stack SaaS coding platform with structural codebase analysis. Multi-provider AI routing (GPT-4o, Groq, Gemini, Claude), real code execution, autonomous repo monitoring, and health scoring.

FastAPI + Next.js + PostgreSQL + Redis + Celery · 11 AI providers

Try the beta →

Live ▬

Compute at Scale

13B+ AI tokens processed across all projects in 48 days. $9,100+ in equivalent API compute, entirely self-funded. Every token tracked and auditable through a custom usage dashboard.

31x the Anthropic Max plan · 48 days · Zero institutional backing

View usage dashboard →

Live ▶

Full-Stack Server Infrastructure

DigitalOcean droplet running all production services simultaneously. FastAPI backends, Cloudflare tunnels, SQLite + PostgreSQL databases, Celery workers, systemd services, UFW firewall, fail2ban, SSH key-only auth.

4 production services · 8GB RAM · Ubuntu 24.04 · NYC3

View all repos →

The Numbers

What building at this scale actually looks like.

13.0B

AI tokens processed

That's the equivalent of reading every English Wikipedia article 65 times. This is how much AI compute Bryan uses building across 4 projects over 48 days.

$9,112

est. compute (API equiv.)

Estimated API-equivalent cost over 48 days based on token counts from local session logs and published Anthropic pricing. 31x the Max plan. Self-funded from Phoenix.

44+

research subjects

Real people who completed a live psychology study. Each one wrote authentic responses that are now part of a peer-reviewable dataset on consciousness and expression.

0.311

Pearson correlation

A statistically significant correlation between how much someone expresses and how comfortable they feel doing it. This number has held steady from 14 subjects to 44.

10B–100B

parameter target

The size of the custom language model being built from scratch. For context, GPT-2 was 1.5B. This targets the scale where real language understanding begins to emerge.

375,000+

traffic snapshots

Continuous monitoring data from 8 Phoenix freeway corridors collected every 10 seconds. Real AZ-511 data from ADOT.

active projects

Running simultaneously: a consciousness study, a language model, a coding platform, and a traffic intelligence system. No team. One person.

institutional backing

No university, no lab, no research grant (yet), no venture capital. Every project is self-funded and self-directed from a home office in Phoenix.

From Phoenix, Arizona. Without permission. Without a lab.

Philosophy

Real work over everything. Not the appearance of progress. The actual thing. If the numbers don't confirm it, the idea doesn't survive.

Shoot for the scale that makes people uncomfortable when you say it out loud. 10 billion parameter models. Peer-reviewed consciousness research. From Phoenix, without permission.

The gap between knowing and saying is the most interesting problem in existence. We are closing it. That is what all of this is.

Contact

Collaboration, research, and impossible ideas.

Email bryanleonard237@gmail.com Phone 480-256-9684 GitHub github.com/TheArtOfSound LinkedIn linkedin.com/in/bryan-leonard-54155218a Preprint zenodo.org/records/19242315 Dashboard Claude Usage Dashboard The Gate theartofsound.github.io/thegate

For Municipal & Transportation Leaders

We built a live predictive cascade model for the I-10 and I-17 corridors that generates 2-minute crew dispatch recommendations. It's running right now on real AZ-511 data.

See the live Phoenix traffic dashboard →

For Researchers & Academics

The EGC framework proposes that expression gates consciousness through a measurable function. The study is still live and the preprint is on Zenodo. Your data would matter.

Take the study or read the preprint →

For Compute & Cloud Providers

We are pushing $9K+ equivalent in API compute and training custom architectures without institutional backing. We are looking for compute grants to scale LOLM to frontier class.

View our compute usage dashboard →

For Engineers & Builders

Everything here was built by two people from a home office in Phoenix. Every repo is public. If you're building something hard and want to collaborate, reach out.

Browse all repos on GitHub →

The study is still live. Your data would matter.

Take the EGC study →